A forest leaf area index inversion acquisition method and system based on data fusion

Abstract

The application discloses a forest leaf area index inversion acquisition method and system based on data fusion. The acquisition method comprises the following steps: performing cloud removal processing and cutting on remote sensing image data of a research area to obtain a segmented image; based on land cover data, obtaining forest pixels in each segmented image, and performing spatial aggregation matching with resolution to obtain aggregated remote sensing images; extracting a vegetation index from the segmented image; obtaining forest canopy height data of the research area; taking reflectivity values of different bands in the aggregated remote sensing images, the vegetation index, the forest canopy height data and a sun sensor related parameter as inputs, and taking the forest leaf area index as output to construct a forest leaf area index fusion prediction model, and obtaining a prediction result. The LAI distribution result obtained by the application has more spatial details, is more suitable for mountain forest ecosystems with large spatial heterogeneity, and can accurately and rapidly monitor mountain vegetation parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing technology application, in particular to the field of obtaining forest leaf area index by remote sensing technology. BACKGROUND

[0002] Forest height, forest leaf area index (LAI) and other forest parameters are indispensable important characteristic parameters in ecological process models and forest carbon cycle models, which can reflect the level of forest net primary productivity and the advantages and disadvantages of forest resources to a certain extent, and are often used in the construction of forest biomass models. Accurate estimation and acquisition of forest leaf area index and the like are beneficial to improving the modeling accuracy of forest aboveground biomass and have important significance for deep understanding of regional and even global climate and environmental change rules.

[0003] At present, forest parameters are mostly obtained by using complex inversion models, which have many input parameters, high field measurement work intensity, and the physical models used are mostly based on flat ground and do not take into account the influence of topographic factors, so the results deviate from the true value and the accuracy is insufficient. Among them, in view of the problems of great difficulty in measurement work, high data acquisition cost and low accuracy, an effective technical means is to fully utilize remote sensing technology, combine remote sensing data with existing methods to improve the estimation accuracy of forest parameters and reduce the difficulty. However, the current LAI remote sensing calculation products such as MODIS LAI and other global land surface parameter products are low-resolution products, which are more suitable for large-scale research, and the medium-resolution LAI products for small-area research do not consider vegetation diversity and regional differences, so the accuracy of the results is insufficient. SUMMARY

[0004] In view of the defects of the prior art, the purpose of the present application is to provide a forest leaf area index inversion and acquisition method which fully considers the topographic factors and scale matching problems and fuses multi-spectral remote sensing data, which can be used in mountainous areas and other regions with large topographic relief and can obtain accurate and high-resolution forest leaf area index distribution.

[0005] The technical scheme of the present application is as follows:

[0006] A forest leaf area index inversion and acquisition method based on data fusion, comprising:

[0007] S1, pre-processing and pixel cutting of remote sensing image data of a region to be obtained forest leaf area index, i.e. a research region, to obtain a segmented image map which eliminates cloud pixels and is segmented into several research plots;

[0008] S2 extracts a mask data of a forest region from land cover data of the study area by a mask extraction method, applies the mask data to the segmented image maps obtained in S1 to obtain forest pixels in each segmented image map, and obtains an aggregated image with improved resolution by spatial aggregation of the obtained forest pixels;

[0009] S3 extracts a vegetation index of the study area from the segmented image maps according to reflection values of different wave bands in the segmented image maps;

[0010] S4 obtains forest canopy height data of the study area through a geographic information system of the study area, and screens out heterogeneous data in the forest canopy height data through a standard deviation and a coefficient of variation of the forest canopy height data to obtain screened canopy height data;

[0011] S5 obtains a forest leaf area index of the study area through a remote sensing data inversion tool;

[0012] S6 constructs and trains a forest leaf area index fusion prediction model to obtain a forest leaf area index inversion estimation value and a forest leaf area index distribution prediction map through the model after training or training and testing; the fusion prediction model is a random forest regression model, which takes reflection values of different wave bands in the aggregated image or the optimized aggregated image, the vegetation index value, the screened canopy height data, and a solar zenith angle, an observation zenith angle and a relative azimuth angle obtained according to a sun sensor as inputs, and takes the forest leaf area index as output.

[0013] According to some preferred embodiments of the present application, the remote sensing image data in S1 is remote sensing image data that has been subjected to radiation calibration and atmospheric correction.

[0014] According to some preferred embodiments of the present application, the segmented image maps are obtained by:

[0015] S11 performs cloud removal on remote sensing image data of the study area that has been subjected to radiation calibration and atmospheric correction through a remote sensing image processing product to obtain cloud-removed image data;

[0016] S12 sets a reflectance threshold of visible light and near-infrared light wave bands in the cloud-removed image data, filters out residual cloud pixels according to the reflectance threshold, and obtains a preprocessed cloud-removed image map;

[0017] S13 crops the preprocessed cloud-removed image map according to the boundary division of different study plots to obtain segmented image maps corresponding to each study plot.

[0018] According to some preferred embodiments of the present application, the reflectance threshold is 0.3.

[0019] According to some preferred embodiments of the present application, S2 comprises:

[0020] S21 performs forest region mask extraction on a 10m resolution land cover dataset, and applies the extracted forest region mask to the segmented image obtained in step S1 to obtain a segmented extraction image, which is an extraction image of only forest region pixels of each segmented image;

[0021] S22 performs spatial aggregation on the segmented extraction image to a resolution of 250m to obtain an aggregated image with improved resolution.

[0022] According to some preferred embodiments of the present application, S2 further comprises:

[0023] The pixels in the obtained aggregated image are matched with the forest leaf area index of the study area by simple averaging method, and the standard deviation and coefficient of variation in the matching are calculated, and the heterogeneous pixels in the aggregated image are screened out according to the coefficient of variation to obtain an optimized aggregated image.

[0024] According to some preferred embodiments of the present application, S2 further comprises:

[0025] The forest region mask obtained in S21 is spatially aggregated to a resolution of 250m to obtain a forest coverage image;

[0026] According to the comparison between the obtained forest coverage image and the aggregated image, the misclassified pixels in the aggregated image are removed to obtain an optimized aggregated image.

[0027] According to some preferred embodiments of the present application, the vegetation index comprises normalized difference vegetation index NDVI, built-up index IBI, difference vegetation index DVI, ratio vegetation index RVI and green chlorophyll index CIg.

[0028] According to some preferred embodiments of the present application, the aggregated image is obtained from a Sentinel-2 remote sensing image, and the reflectance values of different bands in the aggregated image or the optimized aggregated image comprise the reflectance of the green band, the red band, the near-infrared band, the red edge 1 band, the red edge 2 band, the red edge 3 band, the far-infrared 1 band and the far-infrared 2 band of the image at a resolution of 10m and 20m.

[0029] The application further provides a forest leaf area index inversion system, which applies any one of the forest leaf area index inversion acquisition methods and specifically comprises: a data acquisition module for acquiring and / or storing the remote sensing image data, the land cover data, the forest canopy data and the forest leaf area index; a data processing module for performing the preprocessing and pixel clipping in step S1, the screening processes in steps S2, S3 and S4; and a model processing module for performing step S6.

[0030] The application has the following beneficial effects:

[0031] The application can integrate the advantages of various remote sensing data, fuse multispectral data, use a machine learning algorithm, and obtain a forest leaf area index through inversion, and the terrain factor is introduced in the inversion, so that the influence of terrain on forest parameter inversion can be eliminated, and the problem that the forest parameters are difficult to accurately estimate due to terrain can be solved.

[0032] The inversion method of the application can obtain a forest parameter data that has the applicability problem of the medium resolution remote sensing parameter product in the prior art, so that the obtained forest parameter data has more regional particularity.

[0033] The application can have rich forest resources and diversified ecological system regions as research areas, effectively obtain high-resolution leaf area indexes without ground samples, and has higher precision than the inversion model with only two-dimensional optical image data by adding forest canopy structure parameters, and help decision makers to formulate effective forest ecological system environment management policies.

[0034] The application can combine optical data and radar-derived canopy parameters, combine with existing LAI products, establish a machine learning model, utilize the advantages of various different data sources, overcome the data saturation problem of a single data source in existing LAI estimation, and based on the data fusion method, does not need a complex physical model and too many measured vegetation parameters, solves the defect that a large number of field sample points are needed for large-scale LAI estimation, and effectively reduces the operation cost; the application can be directly applied to the Google Earth Engine platform with PB-level geographic spatial analysis capability, can quickly and efficiently process remote sensing images, and generate an LAI distribution map of a research area.

[0035] The leaf area index is estimated by using the data fusion method, the spectral information of multispectral remote sensing data, the solar sensor geometric structure parameters, the vegetation index, the canopy structure parameters and the medium resolution LAI product are combined, the remote sensing data is processed by using the GEE remote sensing big data platform, including: pixel screening, spatial heterogeneity screening, scale matching, sample point selection and finally model establishment, and a 20m resolution leaf area index product of a research area is obtained through inversion.

[0036] The present application contains forest canopy structure parameters which are inversed based on SAR data, and the LAI values are obtained relative to direct LAI inversion results such as GLASS LAI products, the inversion precision is higher, and the spatial resolution is higher. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flow chart of the LAI inversion acquisition method of the present application in the specific embodiment.

[0038] Figure 2 The schematic diagram of the way of pixel aggregation according to the reduceResolution function in Example 1.

[0039] Figure 3 The vegetation index extraction diagram obtained in Example 1.

[0040] Figure 4 The reflectivity distribution diagram of different wave bands in the LAI interval and the total LAI frequency distribution in Example 1.

[0041] Figure 5 The GLASS LAI precision evaluation result comparison of different data sets in Example 1.

[0042] Figure 6 The frequency distribution histogram comparison of the LAI prediction data obtained in Example 1.

[0043] Figure 7 The comparison of the measured LAI and the LAI prediction value (left) or the GLASS LAI value (right) in Example 1.

[0044] Figure 8 The comparison of the GLASS LAI distribution diagram (left) and the LAI distribution diagram obtained by inversion (right) of the research area in Example 1. DETAILED DESCRIPTION

[0045] The present application is described in detail below in conjunction with the embodiments and drawings, but it should be understood that the embodiments and drawings are only used to exemplarily describe the present application, and cannot constitute any limitation on the protection scope of the present application. All reasonable transformations and combinations within the scope of the inventive concept of the present application fall within the protection scope of the present application.

[0046] Referring to the drawings Figure 1 In some specific embodiments, the forest leaf area index inversion acquisition method based on data fusion of the present application comprises the following steps:

[0047] S1: Preprocessing and pixel clipping are performed on remote sensing image data of a region to be obtained forest leaf area index, i.e., a study region, to obtain a segmented image map in which cloud pixels are eliminated and segmented into several study plots.

[0048] In some preferred embodiments, the remote sensing image data uses remote sensing image data that has been radiometrically calibrated and atmospherically corrected, such as remote sensing image data that has been radiometrically calibrated and atmospherically corrected in Google Earth Engine, eliminating the atmospheric influence of the remote sensing image data.

[0049] In some specific embodiments, the preprocessing includes:

[0050] S11: Remote sensing image data of the study region is subjected to cloud removal processing by a remote sensing image processing product such as Sentinel-2L2A to obtain cloud-removed image data;

[0051] In some specific embodiments, the preprocessing includes:

[0052] S12: The obtained cloud-removed image data is subjected to reflectance classification of different spectral bands according to the threshold method in remote sensing image processing, and pixels with a visible and near-infrared (VNIR) band reflectance value greater than 0.3 are screened out to obtain a preprocessed image without residual cloud pixels.

[0053] In some specific embodiments, the pixel clipping includes:

[0054] S13: The cloud-removed image map obtained after preprocessing is clipped according to the boundary division of different study plots to obtain segmented image maps corresponding to each study plot.

[0055] In some specific embodiments, the pixel clipping includes:

[0056] S2: Mask data of forest regions is obtained by extracting the mask data from land cover data of the study region, and the obtained mask data is applied to the segmented image maps obtained in S1 to obtain forest pixels in each segmented image map. The obtained forest pixels are spatially aggregated to obtain an aggregated image with improved resolution.

[0057] In one specific embodiment, step S2 is implemented by the following method:

[0058] (1) Extract forest region mask from ESA's 10m resolution land cover dataset by GEE, the forest region is numbered as 10 in the dataset, apply the extracted forest region mask to the segmented image obtained in step S1 to obtain a segmented extracted image, which only retains the forest region pixels of each segmented image;

[0059] (2) Spatially aggregate the segmented extracted image to a resolution of 250m to obtain the aggregated image with improved resolution; the spatial aggregation can be achieved by the reduceResolution function in GEE, and the pixel weight during aggregation can be set by the degree of overlap between the small pixels in the segmented extracted image and the large pixels in the output aggregated image.

[0060] In some more preferred embodiments, step S2 further comprises:

[0061] (3) Match the pixels in the obtained aggregated image with the LAI data of the study area by simple averaging method, and calculate the standard deviation and coefficient of variation in the matching, and according to the coefficient of variation, exclude the heterogeneous pixels in the aggregated image to obtain the optimized aggregated image.

[0062] In a specific embodiment, the step (3) comprises: matching the pixels in the obtained aggregated image with the GLASS LAI data of the study area by simple averaging method, and calculating the standard deviation and coefficient of variation of each aggregated 250m resolution pixel in the matching, wherein the coefficient of variation is obtained by normalizing the standard deviation by the average value, and is the ratio of the standard deviation to the average value, and the average value represents the spatial reflectivity heterogeneity within an aggregated pixel. If the coefficient of variation of the pixel in any one of the green (560nm), red (665nm) or near-infrared (842nm) spectral bands is greater than 0.1, it indicates that the pixel has high spatial heterogeneity, which may cause a large uncertainty in the relationship between reflectivity and LAI at different spatial scales, and the pixel is discarded.

[0063] In some more preferred embodiments, step S2 further comprises:

[0064] (4) Spatially aggregate the forest region mask with a resolution of 10m extracted in step (1) to a resolution of 250m to obtain a forest cover image;

[0065] Further, in step (3), among the 250m resolution pixels, only when at least 90% of the sub-pixels (i.e., at least 563 out of 625 10m resolution sub-pixels) are classified as forest category, the pixel is considered as a forest pixel;

[0066] (5) According to the comparison between the obtained forest coverage image and the obtained aggregated image, the forest area in the aggregated image that is misclassified is removed.

[0067] S3: Extract the vegetation index from the segmented image obtained in S1 to obtain the value of the vegetation index.

[0068] In some specific embodiments, the vegetation index includes normalized difference vegetation index NDVI, built-up index IBI, difference vegetation index DVI, ratio vegetation index RVI, and green chlorophyll index CIg, which are calculated as follows, respectively:

[0069]

[0070]

[0071] DVI = NIR - R

[0072]

[0073]

[0074] wherein NIR is the reflectance value of the near-infrared band in the image, R is the reflectance value of the red band, in the Sentinel-2 remote sensing image, which corresponds to the pixel values of band 5 and band 4, respectively; SWIR is the reflectance value of the short-wave infrared band, and G is the reflectance value of the green band, which correspond to the pixel values of band 9 and band 3 in the Sentinel-2 remote sensing image, respectively.

[0075] S4: Obtain the forest canopy height data of the study area through the geographic information system of the study area, and screen out the heterogeneous data in the forest canopy height data through the standard deviation and coefficient of variation of the forest canopy height data to obtain the screened canopy height data.

[0076] wherein the forest canopy height data can be extracted by, for example, ArcGIS.

[0077] S5: Obtain the LAI inversion data of the study area through a remote sensing data inversion tool.

[0078] The remote sensing data inversion tool used is, for example, GLASS LAI product.

[0079] S6 constructs and trains a random forest regression model based on the obtained aggregated image, filtered canopy height data, vegetation indices, LAI inversion data, and solar parameters obtained from solar sensors to obtain a fusion prediction model for LAI. This model further obtains LAI inversion estimates and a 20m resolution LAI distribution prediction map. The inputs of the random forest regression model are the reflectance data of the aggregated image, normalized solar zenith angle (SZA), observed zenith angle (VZA), relative azimuth angle (RAA), normalized vegetation index (NDVI), ratio vegetation index (RVI), difference vegetation index (DVI), urban building index (IBI), green light chlorophyll index (CIg), and filtered canopy height data. The output is LAI inversion data.

[0080] In a preferred embodiment, the input to the random forest regression model includes:

[0081] The resulting aggregated images contain reflectance data at 10m and 20m resolutions, including: reflectance in the green band (560nm), red band (665nm), near-infrared band (842nm), red-edge band 1 (705nm), red-edge band 2 (740nm), red-edge band 3 (783nm), far-infrared band 1 (1610nm), and far-infrared band 2 (2190nm); normalized solar zenith angle (SZA), observed zenith angle (VZA), relative azimuth angle (RAA), normalized difference vegetation index (NDVI), ratio vegetation index (RVI), difference vegetation index (DVI), urban building index (IBI), green chlorophyll index (CIg), and filtered canopy height data.

[0082] Considering that the blue band is most sensitive to atmospheric scattering effects, which can easily lead to significant uncertainties in the results, reflectance data of the blue band were not used in the above implementation methods.

[0083] In some specific embodiments, when training the random forest regression model, the RF tree value is set to 100 and the minimum leaf group is set to 5; 80% of the samples are used as the training dataset and 20% are used as the test dataset; the model can be standardized first to fix its value range in the range of [0, 1].

[0084] In some preferred embodiments, during the training process, an LAI frequency distribution map can be plotted based on the LAI inversion data obtained in step S5, and the obtained LAI inversion data can be matched with the reflectance of different spectral bands to obtain a distribution map of LAI values ​​changing with spectral bands. The difference between the model-predicted value and the obtained value can be intuitively observed through the distribution map of LAI values ​​changing with spectral bands and the LAI frequency distribution map.

[0085] In some specific embodiments, it includes:

[0086] (1) The LAI inversion data obtained by S5 is grouped into 8 groups with an interval of 0.5. Samples with reflectance in the green (560nm), red (665nm) and near-infrared (842nm) bands below 1.5 interquartile range (IQR) are removed from each group. That is, samples with reflectance less than Q1-1.5IQR and greater than Q3+1.5IQR in each group are considered outliers. Q1 is the first quartile in each group and Q3 is the third quartile in each group. The filtered LAI samples are obtained.

[0087] (2) Based on the correspondence between LAI values ​​and reflectance in different spectral bands, a distribution map of the LAI samples after screening is plotted to observe the saturation of LAI values ​​in a certain band. Furthermore, during the prediction process, the accuracy of prediction is improved by dividing the data into saturated and unsaturated segments. The corresponding process is as follows: In the screened LAI samples, the LAI value corresponding to the 0-1 group is 0-2, and its corresponding green band reflectance range is 0.10-0.12. Based on this, a distribution map of LAI values ​​changing with spectral bands can be plotted.

[0088] (3) Obtain the LAI frequency distribution map through GLASS LAI to intuitively show the maximum value, minimum value and data distribution trend of LAI in the region, so as to facilitate comparison with the predicted LAI data.

[0089] Based on the above method for obtaining forest leaf area index (FIA) inversion, the present invention can further provide a system for obtaining forest FIA inversion, which includes:

[0090] A data acquisition module for acquiring and / or storing the remote sensing image data, the land cover data, the forest canopy data, and the LAI inversion data; a data processing module for performing the preprocessing and pixel cropping in step S1, and the filtering processes in steps S2, S3, and S4; and a model processing module for performing step S6.

[0091] Through the above system, this invention can overcome the data saturation problem of a single data source in forest leaf area estimation by leveraging the complementary advantages of multiple different data sources through remote sensing data fusion technology; it solves the problem of requiring a large number of field sample points when estimating LAI over a large area, and the data fusion-based method does not require complex physical models or excessive measured vegetation parameters, effectively reducing computational costs; and it can use the Google Earth Engine platform to quickly and efficiently process remote sensing imagery and generate forest leaf area index maps of the study area. The system can also combine canopy structure parameters derived from optical and radar data, perform cloud and snow removal processing on multispectral data in GEE to obtain surface reflectance data, extract solar sensor geometric parameters and vegetation indices corresponding to the optical data, extract forest cover indices for the study area using land cover products, and establish a machine learning model with these data and medium-resolution LAI products, using the data fusion method to invert the leaf area index of the study area.

[0092] Example 1

[0093] The forest leaf area index for a certain region is obtained through the following process:

[0094] (1) Use GEE data preprocessing to filter pixels in the Sennnel-2 remote sensing data of the area to obtain cloud-free remote sensing images;

[0095] (2) Extract forest canopy data of the area using ArcGIS, spatially aggregate the data and put it into the inversion model;

[0096] (3) Forest area mask extraction was performed using the 10m resolution land cover product from ESA. The forest area was numbered 10 in the land cover product. This operation was performed in GEE. The extracted forest area mask was applied to the declouded remote sensing image obtained in step (1) to obtain a remote sensing image that retains only Sentinel-2 pixels of the forest area. During the training sample generation process, the Sentinel-2 image of the forest mask was spatially aggregated to a resolution of 250m using the reduceResolution function in GEE. The pixel weights in the aggregation process were the overlap between the aggregated small pixels and the large pixels specified by the output projection.

[0097] The principle of the reduceResolution function is as follows: Figure 2 As shown.

[0098] (4) A simple averaging method was used to match Sentinel-2 pixels with GLASS LAI data. In addition to the mean, the standard deviation and coefficient of variation for each 250m pixel aggregate were calculated. The coefficient of variation was derived from the standard variation normalized to the mean and is the ratio of the standard deviation to the mean. This mean represents the spatial reflectance heterogeneity within a 250m Sentinel-2 pixel aggregate. If the coefficient of variation for any of the green (560nm), red (665nm), or near-infrared (842nm) spectral bands was greater than 0.1, the 250m pixel was discarded because its high spatial heterogeneity could lead to significant uncertainty in the relationship between reflectance and LAI at different spatial scales.

[0099] (5) Use GEE’s RF algorithm to fuse GLASS LAI and Sentinel-2 data and canopy height data to train the model;

[0100] (6) Construct a model in GEE, with input parameters including vegetation index, surface multi-band reflectivity and corresponding solar sensor geometry, and canopy height;

[0101] (7) Vegetation indices were extracted from Sentinel-2 remote sensing images of Wolong Nature Reserve. Five vegetation indices were selected: Normalized Difference Vegetation Index (NDVI), Index-based Builtup Index (IBI), Difference Vegetation Index (DVI), Ratio Vegetation Index (RVI), and Green Chlorophyll Index. The extraction results are as follows: Figure 3 As shown;

[0102] (8) Utilize GEE data analysis functions to preprocess, scale transform, and extract indices from remote sensing data. Figure 4 );

[0103] (9) Regression analysis was performed on the effective leaf area index measured in the field and Sentinel-2 LAI and GLASS LAI respectively to observe their fit at the data analysis level. Figure 5 , 6 7).

[0104] (10) Utilize GEE's machine learning modeling capabilities to construct a predictive model and perform spatial mapping of the forest LAI ( Figure 8 ).

[0105] The LAI products and remote sensing data used in the above process were collected at the same time as the field data. The main tree species distributed in the study area are coniferous forests.

[0106] in, Figure 4 (ac) represents the reflectance distribution of Sentinel-2 in the green, red, and near-infrared bands across different LAI grouping intervals after outlier removal. Figure 4 As can be seen, the reflectivity of the green and red light bands generally decreases with increasing LAI, while the near-infrared reflectivity shows a relatively gradual increase. Figure 4 (a) and Figure 4 As can be seen in (b), when the number of LAI groups is greater than 5, the LAI value is greater than 2.5, and the reflectivity of Sentinel-2 in the green and red light bands gradually approaches saturation. Figure 4 (c) The reflectance in the near-infrared band does not change significantly. This is related to the leaf structure and tissue of plants. These factors often do not play a significant role in the short-wavelength red and green light. In addition, the LAI value of coniferous forests is generally low and does not easily tend to saturate relative to the vegetation during the growing season. Therefore, the change in the near-infrared band is not obvious. Figure 4 (d) is a histogram of GLASS LAI values ​​for all pixels. The graph shows that LAI follows a normal distribution, with a maximum value of 4.2, a minimum value of 0, a mean of 1.68, and a standard deviation of 0.48. After removing outliers and filtering pixels, 15,635 sample points remained. These filtered points were considered clean pixels, each containing a GLASS LAI value, the average reflectance values ​​of the visible, near-infrared, and short-wave infrared bands of Sentinel-2 data, the geometric parameters of the solar sensor, the forest canopy height, and the extracted vegetation index values.

[0107] Figure 5 The graph shows the results of accuracy evaluation using training data, test data, and all data respectively, with R... 2 All values ​​are above 0.83. The MSE (Mean Sequence Equation) for evaluation and validation using training data is 0.13, the RMSE is 0.23, and the goodness of fit Rfit is [value missing]. 2 The correlation coefficient is 0.83, and the Pearson correlation coefficient is 0.91; the MSE evaluated using the test data is 0.14, the RMSE is 0.22, and the R-squared value is [missing information]. 2 The correlation coefficient was 0.85, and the Pearon correlation coefficient was 0.92; the MSE evaluated using all data was 0.15, the RMSE was 0.24, and the R-squared value was [missing data]. 2The correlation coefficient is 0.83, and the Pearson correlation coefficient is 0.91. It can be seen that the model fit of this method is good, and the high accuracy evaluation results obtained using the data fusion-based method indicate that the method of fusing GLASS LAI, Sentinel-2, and canopy height data can effectively estimate the leaf area index in forest areas. The GEE platform can quickly process the data and perform modeling, and this method does not require field measurement data.

[0108] Figure 6 Histograms of leaf area index (LAI) distributions at 20m in forest areas were generated using the training dataset, test dataset, and all data. All three sets of predicted data follow a normal distribution. Specifically, the predicted LAI using the training data ranged from 0.47 to 2.89, with a mean of 1.71 and a standard deviation of 0.31; the predicted LAI using the training data ranged from 0.5 to 2.84, with a mean of 1.53 and a standard deviation of 0.41; and the predicted LAI using the training data ranged from 0.5 to 2.81, with a mean of 1.61 and a standard deviation of 0.36. Figure 6 histogram distribution and Figure 4 (d) The consistency also indicates that the model has a good predictive effect on leaf area index.

[0109] Figure 7 (a) shows the accuracy evaluation results of the actual values ​​and predicted values. Figure 7 (b) shows the fitting accuracy results between the true values ​​and the GLASS LAI product. Figure 7 It can be seen that the leaf area index measured on the ground using the LAI2000 canopy analyzer and the leaf area index estimated from Sentinel data have a good consistency, with R0.05 2 The R² value is 0.78, the MSE is 0.16, and the fitting accuracy between the LAI2000 measurements and GLASS LAI is essentially the same as the accuracy of the GLASS LAI product itself. 2 The mean squared error (MSE) was 0.74, and the mean squared error (MSE) was 0.17. Without considering the errors of field instruments, the leaf area index estimated using the data fusion method showed slightly higher accuracy in verification with measured LAI data than the data accuracy of the GLASS LAI product.

[0110] Figure 8A comparison is made between the GLASS LAI distribution map of the study area and the 20m LAI distribution map obtained by inversion in this embodiment. It can be seen that the 20m LAI distribution map predicted and plotted on the GEE platform using the data fusion method has a consistent spatial distribution trend with the GLASS LAI product of the same period, but the LAI predicted by Sentinel-2 provides more spatial detail than the GLASS LAI. In both maps, LAI values ​​range from 0 to 4. On the 20m LAI distribution map, it is clearly observable that in the central region of the forest cover area, there are some points with LAI values ​​close to 0, and in the eastern region, there are also some smaller values. Observing the detailed magnified images of the two maps, it can be seen that the LAI in the same area in both maps has the same distribution characteristics, but the 20m LAI map shows more spatial detail in the distribution of leaf area index. The higher-resolution forest LAI map has stronger regional applicability than the medium-resolution LAI product, helping to alleviate the uncertainty caused by spatial heterogeneity in complex terrain.

[0111] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for obtaining forest leaf area index (LAI) based on data fusion, comprising: S1: preprocessing and pixel cropping of remote sensing image data of a study area where the forest LAI is to be obtained, to obtain a segmented image map where cloud pixels are eliminated and segmented into a plurality of study subzones; S2: extracting a forest area mask from land cover data of the study area by a mask extraction method, applying the obtained mask data to the segmented image map obtained in S1, obtaining forest pixels in each segmented image map, and spatially aggregating the obtained forest pixels to obtain an aggregated image with improved resolution; S3: extracting vegetation indices of the study area from the segmented image map according to reflectance values of different bands in the segmented image map; S4: obtaining forest canopy height data of the study area through a geographic information system of the study area, and screening out heterogeneous data in the forest canopy height data through a standard deviation and a coefficient of variation of the forest canopy height data, to obtain screened canopy height data; S5: obtaining the forest LAI of the study area through a remote sensing data inversion tool; S6: constructing and training a forest LAI fusion prediction model, and obtaining forest LAI inversion estimates and a forest LAI distribution prediction map through the trained or trained and tested model; the fusion prediction model is a random forest regression model, which takes reflectance values of different bands in the aggregated image or the optimized aggregated image, the vegetation index value, the screened canopy height data, and a solar zenith angle, an observation zenith angle, and a relative azimuth angle obtained from a solar sensor as inputs, and takes the forest LAI as an output.

2. The forest leaf area index retrieval method according to claim 1, characterized in that, The remote sensing image data in S1 is remote sensing image data that has been subjected to radiation calibration and atmospheric correction.

3. The forest leaf area index retrieval method according to claim 1, characterized in that, The segmented image map is obtained by: S11: performing cloud removal on remote sensing image data of the study area that has been subjected to radiation calibration and atmospheric correction through a remote sensing image processing product, to obtain cloud-removed image data; S12: setting reflectance thresholds for visible and near-infrared light bands in the cloud-removed image data, filtering out residual cloud pixels according to the reflectance thresholds, and obtaining a preprocessed cloud-removed image map; S13: cropping the preprocessed cloud-removed image map according to the boundary division of different study subzones, to obtain segmented image maps corresponding to each study subzone.

4. The forest leaf area index retrieval method according to claim 3, characterized in that, The reflectance threshold is 0.

3.

5. The forest leaf area index retrieval method according to claim 1, characterized in that, S2 comprises: S21: performing forest area mask extraction on a 10m resolution land cover data set, applying the extracted forest area mask to the segmented image map obtained in step S1, and obtaining an extraction image of each segmented image map that only retains forest area pixels, i.e., a segmentation extraction image; S22: spatially aggregating the segmentation extraction image to a resolution of 250m, to obtain the aggregated image with improved resolution.

6. The forest leaf area index retrieval method according to claim 5, characterized in that, S2 further comprises: matching the pixels in the obtained aggregated image with the forest LAI of the study area by a simple average method, calculating the standard deviation and the coefficient of variation in the matching, and screening out heterogeneous pixels in the aggregated image according to the coefficient of variation, to obtain an optimized aggregated image.

7. The forest leaf area index retrieval method according to claim 5, characterized in that, S2 further comprises: spatially aggregate the forest area mask obtained in S21 with a resolution of 10m to a resolution of 250m to obtain a forest coverage image; according to a comparison between the obtained forest coverage image and the aggregated image, eliminate the misclassified pixels in the aggregated image to obtain an optimized aggregated image.

8. The forest leaf area index retrieval method according to claim 1, characterized in that, The vegetation index includes a normalized difference vegetation index NDVI, a built-up index IBI, a difference vegetation index DVI, a ratio vegetation index RVI, and a green chlorophyll index CIg. 9.The forest leaf area index inversion acquisition method according to claim 1, characterized in that, The aggregated image is obtained according to a Sentinel-2 remote sensing image, and the reflectivity values of different bands in the aggregated image or the optimized aggregated image include the reflectivity of the green band, the red band, the near-infrared band, the red edge 1 band, the red edge 2 band, the red edge 3 band, the far-infrared 1 band, and the far-infrared 2 band of the image at a resolution of 10m and a resolution of 20m. 10.The forest leaf area index inversion system using the forest leaf area index inversion acquisition method of any one of claims 1-9, comprising a data acquisition module for acquiring and / or storing the remote sensing image data, the land cover data, the forest canopy data, and the forest leaf area index. a data processing module for performing the preprocessing and pixel clipping in S1, the screening process in S2, S3, and S4; and a model processing module for performing S6. a data processing module for performing the preprocessing and pixel clipping in S1, the screening process in S2, S3, and S4;

Images